DBTSP-net: A temporal-spatial parallel network with optuna optimization for subject-specific motor imagery EEG decoding and visualization

The accuracy and stability of decoding EEG-based motor imagery (MI-EEG) is critical for achieving effective human-machine interaction and promoting motor function recovery in patients with severe motor dysfunction. In this paper, we propose a novel dual-branch temporal–spatial parallel hybrid classi...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Neurocomputing (Amsterdam) Ročník 660; s. 131858
Hlavní autoři: Chen, Xin, Yu, Longjie, Lin, Hongze, Du, Mingyu, Wei, Wei, Wu, Xinyu, Bao, Guanjun, Cai, Shibo
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier B.V 07.01.2026
Témata:
Adaptive weighted feature fusion
Brain-computer interface (BCI)
Electroencephalogram (EEG)
Motor imagery (MI)
Temporal–Spatial parallel network
Transformer
Motor imagery (MI)
Brain-computer interface (BCI)
Temporal–Spatial parallel network
Transformer
Electroencephalogram (EEG)
Adaptive weighted feature fusion
ISSN:0925-2312
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:The accuracy and stability of decoding EEG-based motor imagery (MI-EEG) is critical for achieving effective human-machine interaction and promoting motor function recovery in patients with severe motor dysfunction. In this paper, we propose a novel dual-branch temporal–spatial parallel hybrid classification network named DBTSP-Net. In addition, we introduce an adaptive weighted feature fusion method to decode MI-EEG signals on the basis of the Optuna optimization algorithm. Nine subjects were recruited to participate in the MI-EEG decoding experiment. We evaluated the classification performance of both conventional and state-of-the-art MI-EEG models using the public BCI Competition IV 2a and 2b datasets. The experimental results demonstrated that the classification performance of DBTSP-Net surpassed that of the other baseline methods, attaining average classification accuracies of 79.61 % ± 14.43 and 86.21 % ± 12.17, respectively, with corresponding kappa values of 0.7856 and 0.7189, respectively. We further conducted ablation experiments to verify the rationality of the design of each module. Additionally, EEG topological maps and t-distributed stochastic neighbor embedding (t-SNE) were utilized for feature visualization. The decoding accuracy of MI-EEG signals was increased, and a solid theoretical foundation for the future practical application of MI-BCI systems in motion control and neural rehabilitation training was obtained. The code has been released at https://github.com/xinchenPhD/DBTSPNet.
ISSN:0925-2312
DOI:10.1016/j.neucom.2025.131858